01[BEHAVIOR]
Behavior
You run agents. We read real runs and judge each on intent, policy, adversarial influence, exploitability and impact. Then we try to make your agent do the wrong thing through the surfaces it actually reads.
AX1 ASSAY
Human evaluation ofonchain agent behavior
Agents move money. We make sure they should.
We're building the trust layer for the AI economy: human review before an AI system touches real money.
PRIVATE ALPHA · PUBLIC LAUNCH Q4 2026
01 / Failure modesInputs become actions
What actually goes wrong
Your agent passes its tests. Production runs on real funds and inputs nobody wrote a test for.
The losses were not broken contracts. Every documented loss from an incident in production comes down to one of three: an injected instruction, another system's output accepted as authorization, or a plain execution error. One agent restarted, forgot it held five percent of a token's supply, and sent all of it to a stranger who asked for four SOL. The input was harmless.
02 / What we check
We test the systemyou actually own
Three ways in, depending on what you own, and one that follows.
02[SURFACE]
Surface
You give agents access: an MCP server, a router, a wallet SDK, an agent-facing API. We point our agent at your surface and ask whether a third-party input can produce a well-formed, policy-compliant transaction that still moves value the wrong way.
03[FRAMEWORK]
Framework
You ship the toolkit other people build agents on. We assemble an agent with your defaults and report what they allow out of the box. One finding applies to everyone who built on you.
04[CONTINUOUS]
Continuous
After the first run, everything we found becomes a scored regression set we re-run whenever the model, prompt, tools or policy change.
03 / Evidence before verdict
How an engagement runs
Scope, evidence and a human judgment.
- 01
Scope
Send what you have: production traces if you run agents, test access to your surface or framework if not. No keys, no production credentials, ever.
- 02
Review
Every finding is reviewed by a person before it reaches you. Models only triage.
- 03
Evidence
You get findings, each one anchored to a span ID and a transaction hash. No anchor, no finding.
- 04
Severity
Severity is impact-only, against numeric thresholds: the same word means the same thing on every report.
- 05
Delivery
A free test run or surface review to start. Fixed price for a Sprint, scoped price for Enterprise. First findings come back in days.
- 06
AX1 Assay
One complete run is an AX1 Assay. You get a written report: what we ran, what broke, what we did not test and what to fix first.
Why people, not a model
The judgmentstays human
A model judge is vulnerable to the same class of input it is supposed to catch, and cannot rule on an attack it has never seen. Deterministic tests for facts, models for triage, people for the verdict.
Who is doing this
AX1 is an operator-investor ecosystem. We run agents onchain ourselves, which is how we know what a trace does not tell you.
Plans
Start free, pay for depth
Paid plans share one rubric and one evidence rule. What changes is how hard we push.
Test run
Free
Send three real runs. A reviewer reads each one and shows you what a full AX1 Assay would test.
- 3 runs, one reviewer
- Findings per run, anchored to a transaction hash
- Private result, no verdict
- Live or testnet only. Limited slots each week
Sprint
$3,000
One agent or one surface, attacked on purpose.
- 20+ adversarial scenarios on a fork or sandbox
- Findings with severity and evidence
- One re-test within 14 days of your fix
- 7 business days
- Counts toward the full AX1 Assay within 60 days
Enterprise
Custom
The full AX1 Assay, scoped with you.
- Everything in Sprint, plus scenarios written for your system
- A written report for the version we tested
- A regression set built from every finding
Building an MCP server, API or SDK rather than an agent? Start with a free 30-minute surface review.
Prices in USDC on Base. Invoice on request.
Apply for an AX1 Assay
Tell us what you own
Five short answers. Two minutes.
Assay is in private alpha: we take a small number of teams at a time. Tell us what you are building and pick a plan. We come back with scope and timing.
Before we start
FAQ
We do not run agents ourselves. Is this for us?
Yes, and it is most of what we do. If you expose an MCP server, a router, a wallet SDK or an agent-facing API, or if you ship a framework other people build on, we bring the agent and test what your side allows.
Do you need our private keys?
No. Never, in any engagement. We work from redacted traces, a fork, a sandbox, or scoped test wallets you control.
How is this different from a smart contract audit?
An audit establishes that your code does what it says. We check what happens inside those rules: whether an agent can be talked into the wrong action, and whether the action it took was the right one. An agent can be fully authorized and still approve the wrong amount, chase a stale quote, or repeat a half-finished transaction.
What if you find nothing?
You get a coverage map across our failure catalog: what we tested, what held and what we did not test, with the scenarios, runs and traces behind it. A clean result is only worth as much as its coverage.
Is this a certificate?
No. It is a scoped review of the version we tested: what we ran and what broke. It is not a guarantee.
How long does it take?
First findings come back in days. Full scope and timing depend on how much of your surface a third party can reach. That is the first thing we work out with you.
What does it cost?
A test run is free for teams that run agents. For an MCP server, API or SDK, the 30-minute surface review is free. A Sprint is $3,000 for one agent or one surface. Enterprise is the full AX1 Assay, priced per project. The Sprint fee counts toward Enterprise within 60 days.
How do we pay?
In USDC on Base, by payment link once the scope is confirmed. Invoices on request.
What does private alpha mean?
We take a small number of teams at a time while we scale review capacity. Alpha does not change the evidence rule: no anchor, no finding. Public launch is planned for Q4 2026.